Experimental Characterization of Value Locality

A number of recent publications have described various value prediction techniques for exploiting value locality in the data-flow portion of a processor to improve program performance. An examination of the value usage characteristics of programs is appropriate and relevant because it is within the value space that value prediction mechanisms operate. In this paper, we perform an extensive experimental study on value usage patterns of the SPECint95 benchmark programs, and find that they make very sparse and uneven use of the available value space. Furthermore, we suggest a metric for characterizing the value locality of programs and their data sets. This metric, the Value Repetition (VR) factor, represents the redundant redefinitions performed by a program and can be used to formulate a new theoretical upper bound on the speedup possible with per-instruction history-based value prediction techniques.